The Field Missing From the Footnotes

Working Theory: Colin Willis on carrying ideas across disciplines. Beyond the Prompt: Why AI Work Runs on I-O Psychology, Part 1

Economists at Anthropic, the AI company that makes Claude, have published the most detailed account I’ve seen of how AI changes work inside a company that has reorganized around AI. As an industrial-organizational (I-O) psychologist, I read the report as a casebook of my own field’s ideas: a performance review written for a team of AI agents, thousands of standing instructions telling agents what good work looks like, and hiring criteria shifting toward judgment. None of the ten works in the report’s list of works cited comes from I-O psychology. My field argued about its role in a new kind of workplace expertise once before, a decade ago. That expertise was data science and people analytics, the use of workforce data to guide decisions about people. I-O psychology has a second chance now. I think the field gets a seat at the table where AI work is designed only if I-O psychologists learn the language of the people building AI systems.

Every organization adopting AI is asking what happens to people’s jobs when AI takes on more of the work. In October, four economists at Anthropic, Nathan Wilmers, Regina Catipon, Maxim Massenkoff, and Peter McCrory, published one company’s answer (Wilmers et al.). The authors drew on the company’s internal data and on interviews with 17 employees, and they describe Anthropic as a company that “could be a leading indicator of how AI will shift work organization,” an early signal of changes other companies will see later. At Anthropic, much of the work is now done by AI agents: AI models that complete multistep tasks on their own, using tools such as code editors and browsers. By August 2026, the agents Anthropic’s employees directed were logging about six times as many hours as the employees themselves. Someone has to direct and review all of those agent hours.

The finding I kept returning to was about the people reviewing all that work. “Amid this model-driven acceleration,” the authors write, “human attention is a constant bottleneck.” Anthropic’s Claude sessions “amount to around 80% idly waiting for human input”: for most of each session, the agent is waiting for a person to respond. Reading everything those sessions produce would take the company’s non-technical staff more than half their work week. One engineer described facing “an army of Claudes, with every one asking for my attention.” I read the 80% figure as a sign of how much Anthropic’s employees have automated: most of the work outside human review now runs on its own, so quickly and so widely that people have trouble keeping pace with the review. An employee’s job becomes responding to many agents as quickly as possible, and a person can check only so many answers while the number of agents keeps growing.

Anthropic’s employees responded to the flood of agent work by working on the work itself: building tools to see what the agents were doing and documenting how the work should be done. This work on the work, which the authors call meta-work, rose from 10% of the average employee’s working sessions with Claude in fall 2025 to 22% in spring 2026, and “written standing instructions for agents grew from a few hundred in July 2025 to 8,500 a year later.” Those instructions tell agents what good work looks like, a job I return to below.

The authors are careful about what one company can show. Anthropic makes the AI tools the report studied, and employees chose when to start using those tools. The report also measures quality only for code, and the authors write that “there are many parts of Anthropic’s employment practices that cannot be replicated across the economy.” I read the report as evidence from one firm, which is how the authors present their findings.

The report also read like an account of my own past year. On about the same timeline, I began putting more AI agents to work and automating technical work I used to do myself, and I was struck enough by how much one person could now build and run that I left a stable corporate job to build Corymb. The report finds Anthropic’s hiring managers writing more often about autonomy and judgment, and those are the qualities I found myself leaning on. I had to leave my old role to use those qualities fully, because an organization changes more slowly than a person can. Corymb now runs the way the report describes Anthropic running: our meetings are open and transcribed, we communicate in public channels, and AI and human workers see the same information, so information silos don’t form, and the systems we build later don’t automate those silos into place.

My field, under other names

What I kept recognizing, page after page, was my own field. SIOP, the field’s professional society, describes I-O psychology as “the scientific study of working and the application of that science to workplace issues facing individuals, organizations, and teams” (SIOP). The report never uses the name I-O psychology, but the practices it describes are the field’s core methods, adapted for teams that include agents.

The people at Anthropic building these ways of working are doing I-O psychology without the name. I don’t mean that anyone at Anthropic needed a textbook to write a performance review for a set of agents. I mean that the problems AI work raises are the problems my field studies, and the solutions people reach for are ones the field would recognize. Prakriti Singh, writing in SIOP’s practitioner magazine about I-O careers outside traditional roles, makes a similar observation: the field “travels through functions that may use I-O principles explicitly, implicitly, or under another name” (Singh).

A fair objection is that I’m claiming credit for good management. The same practices map onto management science and organizational behavior, and a discipline that stakes a claim to them can sound like a field guarding territory. There’s a real case for leaving the practices unnamed: the practices are working, and the people using them don’t need a label. Naming the practices has costs too: jargon, and a field defending its title instead of its reasoning. Singh’s line about careers applies here: titles are “imperfect proxies for practice.”

My answer is that the principles of I-O psychology are what drive the success, whatever anyone calls them, and the name is how you find the rest of the field’s principles. A discipline connects an idea to its neighbors, which psychologists call a nomological network: the web linking a concept to related concepts and to what can be observed about them (Cronbach and Meehl). A discipline tells you where else to look. A discipline also gives you a way to test ideas and accumulate evidence through practice, along with a community of people working on the same problems. Without the name, practitioners reinvent what’s known and repeat what’s been tried.

Hiring is the clearest example of what organizations lose when they work without the field’s evidence. The report found hiring managers writing more about autonomy and judgment and less about domain expertise. As one manager explained, “I expect models to continue to be jagged, so for my team I want them to be fungible across domains.” The models are uneven, strong at some tasks and weak at others, so the manager wants people who can move between kinds of work. Hiring for fungibility is a reasonable response to roles that change from month to month, and the approach runs straight into one of the oldest problems in my field. I-O psychologists call the difficulty of defining and measuring job success the criterion problem (Austin and Villanova). “Judgment” is hard to define and harder to measure. When roles change faster than anyone studies them, I think the risk is high that organizations fall back on old habits, like hiring for a degree or a company’s name on a résumé. Handing the screening to AI doesn’t remove that risk. In a 2024 audit of language models used to rank résumés against job descriptions, the models favored résumés with White-associated names over Black-associated names in 85.1% of tests. Résumés with female-associated names were favored over male-associated names in only 11.1% (Wilson and Caliskan). Without a careful definition of what the job requires, the model’s preferences stand in for the job’s requirements.

The seat I-O debated once before

If my field has that much to offer, where is I-O psychology in the report? The report’s list of works cited has ten entries: economics journal articles and working papers, a Harvard Business School working paper, and Anthropic’s own studies. None comes from I-O psychology, and the text never mentions psychology. The authors are economists studying labor demand, so I don’t fault their bibliography. I read the missing citations as a sign of who is in the conversation about AI and work, and my field isn’t part of that conversation yet.

I-O psychologists have been here before. In 2015, the SIOP journal Industrial and Organizational Psychology ran a focal article on what big data meant for the field (Guzzo et al.). A focal article is a lead paper the journal invites the field to answer in published commentaries, and the commentaries that followed debated where I-O belonged in big data. One group of commentators urged I-O psychologists to “join business analysts, data scientists, statisticians, mathematicians, and economists in creating the vanguard of expertise.” They also warned that the field needed to “rapidly upskill their capabilities,” since machine learning was “typically not part of the curriculum in I-O psychology programs” (Ducey et al.). Two years later, William Schiemann and Dave Ulrich called analytics “the single biggest area for I-O impact—and visibility” (Schiemann and Ulrich).

Despite those calls, data science and people analytics grew into fields of their own. Practitioners interviewed for SIOP’s magazine describe people analytics as “an interdisciplinary and cross-functional team sport,” one that I-O shares with disciplines such as labor economics and sociology. Rick Guzzo, the lead author of the 2015 focal article, was still arguing in 2023 that “most PA [people analytics] teams should be led by I-Os” (Thoebes and Roesler). Redesigning work around AI is a second chance at the same seat. The lesson from the data-science debate is that expertise about people isn’t enough on its own. I-O psychologists also have to speak the language of the people building the systems.

Two seats at the table

I see two seats open to I-O psychologists, at the table the field missed last time. In the first, the I-O psychologist advises an organization on AI work: how to break the work down for agents, how to write the criteria that manage their performance, and how to open the culture so people and agents work from the same information.

Go back to the engineer facing an army of Claudes. When attention is the bottleneck, AI work leaves people two jobs. The first is deciding what good looks like, which is the criterion problem in a new setting. Anthropic’s 8,500 standing instructions are, in large part, written definitions of what good work looks like for agents. The second is deciding which parts of the work need judgment at all and which should run on fixed rules, a question my field has studied for decades as clinical versus mechanical prediction: whether an expert’s judgment or a fixed formula predicts outcomes better (Grove et al.). Across studies, the formula is right somewhat more often than the expert. The stronger the AI gets at a task, the less a person adds by checking its work, and the more they add by deciding what the work is for.

In the second seat, the I-O psychologist builds. Building takes fluency in how AI systems work. A builder knows that an agent can take on and automate tasks, and that agents can check one another’s work. In a loop, a draft goes back until it meets the criteria; in an adversarial review, one agent’s job is to find faults in another’s work. A builder also knows that workflows can be arranged as graphs of steps, and that tools and rules can be built around a model to improve the model’s performance, which engineers call a harness. The more an I-O psychologist knows about AI, the more of the field they can apply as AI changes work. At Corymb, we work in both seats at once: we rethink how work gets done, in our own business and on client projects, do the work, evaluate the results, and package what works into products or new ways to deliver services.

The second seat changes the field, too. Singh uses a name for that change, domain entanglement, which she first heard from her mentor, Shonna Waters: when an I-O psychologist works deeply inside another domain, “your training shapes how you see problems in the new field, and the field reshapes how you practice I-O.” Singh offers domain entanglement as a framing to react to, not a validated theory, and the phrase describes what I want for my field. I-O can claim these practices only by changing to keep up with them.

Most people meet AI in the chat window, in what I call turn-based prompting: a person writes a prompt, reads the reply, and decides what to ask next. Getting better at each turn is a skill most people are still building, so the rest of this series starts there and climbs toward systems that check and improve their own work. Each article works through the two jobs in the language of the people building the systems, pairing one AI practice with the psychology that explains or improves that practice.

For I-O psychologists, the next step is learning the vocabulary of agents, loops, graphs, and harnesses, and taking on AI work even when the work doesn’t look like I-O. HR leaders can treat AI transformation as work design and study a role before changing what they hire for. Anyone working where AI meets the work itself can ask what the field already knows before treating a problem as new. What happens when you carry an idea from one discipline into another? In this series, the ideas travel both ways: I-O’s ideas are already at work in AI under other names, and the field will change as it takes them back. The seat in this conversation will go to people who can speak both languages, I-O psychology’s and AI’s, and I want my field among them.

The series ahead

Sources

Austin, J. T., & Villanova, P. (1992). The criterion problem: 1917–1992. Journal of Applied Psychology, 77(6), 836–874. https://doi.org/10.1037/0021-9010.77.6.836

Cronbach, L. J., & Meehl, P. E. (1955). Construct validity in psychological tests. Psychological Bulletin, 52(4), 281–302. https://doi.org/10.1037/h0040957

Ducey, A. J., Guenole, N., Weiner, S. P., Herleman, H. A., Gibby, R. E., & Delany, T. (2015). I-Os in the vanguard of big data analytics and privacy. Industrial and Organizational Psychology, 8(4), 555–563. https://doi.org/10.1017/iop.2015.83

Grove, W. M., Zald, D. H., Lebow, B. S., Snitz, B. E., & Nelson, C. (2000). Clinical versus mechanical prediction: A meta-analysis. Psychological Assessment, 12(1), 19–30. https://doi.org/10.1037/1040-3590.12.1.19

Guzzo, R. A., Fink, A. A., King, E., Tonidandel, S., & Landis, R. S. (2015). Big data recommendations for industrial–organizational psychology. Industrial and Organizational Psychology, 8(4), 491–508. https://doi.org/10.1017/iop.2015.40

Schiemann, W. A., & Ulrich, D. (2017). Rise of HR—New mandates for I-O. Industrial and Organizational Psychology, 10(1), 3–25. https://doi.org/10.1017/iop.2016.76

Singh, P. (2026). Domain entanglement: The phrase I needed for the jobs that weren’t on the list. The Industrial-Organizational Psychologist, 64(2). https://www.siop.org/tip-article/domain-entanglement-the-phrase-i-needed-for-the-jobs-that-werent-on-the-list/

Society for Industrial and Organizational Psychology. (n.d.). Industrial and organizational psychology: Science for a smarter workplace [Brochure]. https://www.siop.org/Portals/84/docs/swam/siopgeneralbrochure.pdf

Thoebes, G. P., & Roesler, A. (2023). I-Os and people analytics: Insights from experts in the field. The Industrial-Organizational Psychologist, 61(2). https://www.siop.org/tip-article/ios-and-people-analytics/

Wilmers, N., Catipon, R., Massenkoff, M., & McCrory, P. (2026). How employment is changing at Anthropic. The Anthropic Institute. https://www.anthropic.com/research/employment-at-anthropic

Wilson, K., & Caliskan, A. (2024). Gender, race, and intersectional bias in resume screening via language model retrieval. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 7(1), 1578–1590. https://doi.org/10.1609/aies.v7i1.31748